Fine-tuned unsloth/gemma-4-E4B-it on a curated, native-heavy German SFT corpus to improve general German skill (not a single benchmark). SFT · LoRA · bf16 · reasoning OFF.

Variant: v2

Compute

machine: 1x NVIDIA H100 80GB (a3-highgpu-1g, GCP us-central1-b)
training wall-time: 101.5 min (1.0 epoch(s))
final train_loss: 1.7187

Training parameters

base: unsloth/gemma-4-E4B-it · quant bf16 (load_in_4bit=False, no QLoRA)
lora_r 128 · alpha 256 · dropout 0.0 · target language+attn+mlp (vision frozen)
lr 5e-05 · adamw_8bit · linear · warmup_steps 100
per_device_batch 16 · grad_accum 1 · eff_batch 16 · max_seq 2048 · seed 3407
chat_template gemma-4 (non-thinking) · train_on_responses_only

Data

corpus: peerbench/gemma4-german-sft-corpus (split gemma4-de-v2) — 25756 rows
native-heavy German + high-quality MT + balanced m_arc MCQ; per-source 20% cap; lang-ID + dedup; slimorca / ultra_chat excluded (noise / verbosity).

Logging

W&B gaa-mdk (entity forest-wb), run gemma-4-E4B-v2__sftmix-r128__20260610190427 · date_utc 20260610190427

peerbench/gemma-4-E4B-20260610190427-v2

gemma-4-E4B — German SFT — gemma4-de-v2 (20260610190427 UTC)

Compute

Training parameters

Data

Logging

peerbench/
gemma-4-E4B-20260610190427-v2